Abstract
Scene flow estimation is critical for autonomous perception, yet traditional coarse-to-fine strategies often suffer from error propagation. IHNet (Iterative Hierarchical Network) addresses this by introducing a recursive feedback mechanism. By feeding high-resolution flow and features from previous iterations back into coarse layers, the model self-corrects. Combined with a resampling scheme that aligns source and target point clouds, IHNet significantly outperforms existing methods, achieving a 56% reduction in End Point Error (EPE3D) on the KITTI dataset.
Introduction
Scene flow estimation predicts the 3D motion of points between consecutive frames, a vital task for understanding dynamic environments in autonomous driving. While coarse-to-fine architectures are standard, they possess a "weakest link" vulnerability: errors in initial coarse estimates propagate through higher resolutions. Additionally, the irregular and sparse nature of point clouds makes finding exact correspondences between frames challenging. IHNet addresses these issues by breaking the traditional linear pipeline with iterative refinement and spatial alignment.
Methodology
IHNet introduces two primary innovations to the scene flow pipeline:
- Iterative Hierarchical Guidance: Instead of a single pass, IHNet runs multiple iterations. It takes the high-resolution flow and latent features estimated at the end of one iteration, downsamples them, and feeds them back into the coarse layers of the next iteration. This "top-down" feedback ensures that the coarse layers have a much more accurate starting point for their matching search.
- Resampling Scheme: To solve the correspondence problem, the model uses the flow estimated in the previous iteration to "warp" the source points toward their predicted positions in the next frame. It then finds the nearest neighbors in the actual target frame to create a "resampled" target point cloud. This effectively aligns the two point clouds in space before the next round of estimation, making the matching process significantly easier for the network.
Experimental Results
The model was trained on the synthetic FlyingThings3D dataset and evaluated for generalization on the real-world KITTI Scene Flow benchmark. IHNet sets a new state-of-the-art (SOTA) across multiple metrics:
- Accuracy: On FlyingThings3D, it improved the EPE3D by 20.1% over previous SOTA methods.
- Generalization: On the real-world KITTI dataset, the improvement was even more dramatic, reducing the error by 56.0% (from 0.0277m down to 0.0122m).
- Robustness: The "Acc3D Strict" metric (percentage of points with very low error) reached 97.79% on KITTI, suggesting high reliability in real-world scenarios.
Key Insights and Discussion
The core advantage of IHNet is the observation that "high-res hindsight is 20/20." A downsampled version of a high-resolution estimate is significantly more accurate than an original coarse-level estimate starting from zero. Furthermore, the paper demonstrates that passing back latent features (context) rather than just motion (flow) provides a 19% improvement in accuracy.
While IHNet is highly accurate, the iterative nature comes with a computational trade-off. A 3-iteration pass takes approximately 130.9ms compared to the 70ms baseline. However, the performance gains—particularly the massive reduction in real-world error—make it a compelling choice for robust 3D perception. Future work may explore more efficient ways to inject high-resolution information without full re-encoding of point clouds.
Summary updated